基于MHI的三重运动表现人体动作图像识别方法  

A human motion recognition method of triplet motion representation images based on MHI

作　　者：曹静 龚俊杰[1] 山下阳太郎 TAN Joo Kooi CAO Jing;GONG Junjie;YAMASHITA Youtaro;TAN Joo Kooi(School of Mechanical Engineering,Yangzhou University,Yangzhou 225117,China;School of Engineering,Kyushu Institute of Technology,Kitakyushu 804-8550,Japan)

机构地区：[1]扬州大学机械工程学院,江苏扬州225127 [2]九州工业大学,日本北九州804-8550

出　　处：《扬州大学学报（自然科学版）》2020年第4期48-53,共6页Journal of Yangzhou University：Natural Science Edition

基　　金：国家自然科学基金资助项目(51275447)。

摘　　要：针对传统的老年人智能监护系统中运动历史图像(motion history image, MHI)动作识别方法仅考虑时间序列特征而难以表示相机光轴方向(即深度方向)上的运动问题,基于传统MHI进行改进,提出一种三重运动表现图像(triplet motion representation images, TMRI)识别方法.该方法首先通过跟踪移动物体表面的特征点计算光流和检测消失点得到动作的深度信息,然后创建包含与传统MHI相同的时间序列图像、表现前景出现频率的重叠度图像以及表现相机光轴方向运动信息的深度图像3种类型的运动历史图像来描述目标对象的动作,最后利用各图像的特征对动作进行识别.结果表明基于MHI的三重运动表现图像方法的平均动作识别准确率可达84.35%,明显高于传统MHI方法的识别准确率(55.43%).The traditional motion history image(MHI)method in the elderly intelligent monitoring system is difficult to repesent the motion in the direction of the optical axis of the camera,that is,in the depth direction.Aiming at the difficulty,a motion performance and recognition method called triplet motion representation images(TMRI)which extends MHI is proposed in the paper.On the basis of traditional MHI,two historical images are innovatively proposed to perform and recognize actions.The method calculates the optical flow by tracking the feature points on the moving object and detects the vanishing point,thereby obtaining the depth information of the action.Then creates three types of motion history images(time series image:the same as MHI,overlap image:the frequency of the foreground,depth image:the movement information of the camera’s optical axis)to describe the action of the object.Finally,the features of each image are used to recognize the action.The experimental results show that the average action recognition accuracy of the TMRIs reaches 84.35%,which is significantly higher than that of the MHI method(55.43%).

关 键 词：老年人智能监护 动作表现 动作识别 运动历史图像 三重运动表现 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]